Few languages have as wide a gap between how people talk and how they write. In conversation میخواهم becomes میخوام, است becomes ـه, را shrinks to رو or just ـو, and نان is نون. A transcript has to pick a side, and it does not always pick the same one: careful speech such as news, lectures and audiobooks tends to come out in formal written forms, while casual talk is usually written the colloquial way it was said, sometimes with a formal form slipping in.
Neither is wrong, but it affects what you do next. For subtitles the colloquial form is what viewers expect. For a report or a thesis you will want to bring the text up to written standard, and that is an editing job for a person. The transcript is not a Finglish text either: Latin-letter Persian is not produced.
Persian script leaves short vowels out, so the reader supplies them. That includes the ezafe, the unwritten -e that links کتابِ من, which is heard clearly in speech and mostly invisible on the page. The transcript follows normal practice and does not mark it, except where a ی is conventionally written after a vowel.
The half-space (نیمفاصله) in words like میروم and کتابها is the detail that editors notice first. The transcript may use it, use a full space or join the parts, and it is not consistent throughout a long file. If the text is going to print, run it through a Persian text normaliser; the same step will unify ی and ک, which appear in both Persian and Arabic forms in many fonts and keyboards.
The models have heard far more Iranian Persian, and Tehran speech in particular, than anything else, and that is where results are best. Dari from Afghanistan is the same language in the same script, with different vowels and a good deal of different vocabulary. It is transcribed, but with more errors, and the spelling tends to drift towards Iranian usage. Hazaragi and strong regional accents inside Iran, such as Isfahani, Mashhadi or Yazdi speech at full speed, are harder again.
Tajik is a separate case because it is written in Cyrillic. A Tajik recording is likely to come back in Persian script or as a poor mixture, so this page is not the right tool when you need Tajik Cyrillic text. Kurdish, Pashto, Azerbaijani and Balochi are different languages and are not covered here.
Persian uploads are often long voice messages from family abroad, university classes, interviews for research and journalism, podcasts and YouTube talk programmes. Poetry recitation is transcribed, though classical vocabulary produces more mistakes than everyday speech, and sung poetry over music does not work well. Guests get 10 transcriptions of up to 3 hours each, with 5 hours in total. A free preview opens without an account, and a free account keeps the full transcript. AI chat, summaries and PDF or Word export are part of the paid plans.
The audio is processed on servers we operate, with our own speech models, and is not sent to a third-party speech API. For personal messages and interviews with people who asked not to be identified, that is one less party handling the recording.